OpenAI Prompt Optimizer Is Shutting Down: 7 Alternatives Compared (2026)
Quick answer: OpenAI is retiring its dataset-backed prompt optimizer together with the Evals platform. Evals went read-only for existing users on October 31, 2026, and the dashboard, API, and graders shut down on November 30, 2026. Reusable prompt objects (/v1/prompts) close the same day. OpenAI's own replacement for the scoring half is Promptfoo, which is open source and now part of OpenAI. For the automatic rewriting half, use DSPy if your prompts live in code, Vertex AI's data-driven optimizer if you are on Gemini, or a hosted rewriter like Prompt Builder (free plan, paid plans from $9 a month) if you write prompts by hand for ChatGPT, Claude, or Gemini.
This guide covers what exactly is closing, what the optimizer did, and 7 alternatives sorted by the job you used it for. Prices and docs were checked on vendor pages in November 2026.
What is shutting down, and when
OpenAI announced the changes on June 3, 2026. Its deprecations page and the prompt optimizer guide give this timeline:
| Date | What happens | OpenAI's recommended replacement |
|---|---|---|
| June 3, 2026 | Deprecation announced for the Evals platform, reusable prompt objects, and Agent Builder | |
| October 31, 2026 | Evals become read-only for existing users | |
| November 30, 2026 | Evals dashboard, API, and graders for eval workflows shut down; the dataset-backed prompt optimizer goes with them | Promptfoo |
| November 30, 2026 | /v1/prompts and reusable prompt objects shut down |
Move prompt content into your application code |
| November 30, 2026 | Agent Builder shuts down | Agents SDK or ChatGPT Workspace Agents |
Two details matter if you are migrating this month.
First, OpenAI's Promptfoo migration guide says you recreate each evaluation by hand in a promptfooconfig.yaml file. It "does not require an OpenAI Evals export feature", which also tells you there is no one-click export. Copy your test rows, grader logic, and best prompts out of the dashboard while you can still read them.
Second, if you stored prompts as reusable prompt objects and called them by ID (prompt: { id: "pmpt_123", version: "1" }), those calls stop working on November 30. OpenAI's guide says to move the prompt text into your own code, where you control review, testing, and versioning.
What the OpenAI prompt optimizer actually did
It helps to know which part you are replacing. The optimizer was a chat interface in the platform dashboard, and it was strongest when paired with a dataset. The workflow was:
- Build a dataset with the prompt and at least three rows of responses.
- Give each row at least one grader result or human annotation: a Good/Bad rating, custom annotation columns, or written feedback in an
output_feedbackcolumn. - Click Optimize. A background run rewrites the prompt against that feedback and opens the result in a new tab.
- Test the new prompt, annotate again, and repeat.
OpenAI's docs were honest about the limits: the result is only as good as your graders, and an optimized prompt can still do worse than the original on specific inputs.
So the tool did three jobs, and most replacements only do one or two of them:
- Rewrite: turn a prompt into a better-structured one.
- Score: run the prompt over test cases and grade the outputs.
- Store: keep the prompt somewhere you can call or reuse it.
The 7 alternatives at a glance
| Tool | Replaces | Price | Needs a dataset | Needs code |
|---|---|---|---|---|
| Prompt Builder | Rewrite + store, for hand-written chat prompts | Free plan (5 prompts), then $9, $19, or $49 a month | No | No |
| Promptfoo | Score (OpenAI's official migration path) | Free and open source (MIT); Enterprise is custom | Yes, as test cases | Config file + CLI |
| DSPy | Rewrite against a metric (closest to dataset-backed optimization) | Free, open source (MIT); you pay model usage | Yes, training examples | Python |
| Vertex AI Prompt Optimizer | Rewrite, zero-shot or data-driven, for Gemini | No separate fee listed; Google Cloud model usage | Optional (data-driven mode) | SDK or console |
| Braintrust | Score + AI-assisted iteration (Loop) | Starter $0; Pro $249 a month | Yes | Light |
| PromptLayer | Store (prompt registry) + evals | Free (10 prompts); Pro $49 a month; Team $500 a month | Optional | Light |
| Prompt Optimizer (open source) | Rewrite + side-by-side compare | Free (AGPL-3.0); you pay model usage | No | No (web or desktop app) |
1. Prompt Builder
Prompt Builder is a prompt generator, optimizer, and prompt library for ChatGPT, Claude, Gemini, Grok, DeepSeek, Llama, Mistral, Perplexity, and Cohere. It replaces the casual use of OpenAI's optimizer: you paste a prompt you already use, it asks a few clarifying questions, and it rebuilds the prompt with a clear goal, context, constraints, and output format for the model you picked. Each refine is saved, so you can compare against where you started.
It also covers the "store" job for people who type prompts into a chat window rather than call them from code. The Library holds your prompts with tags and folders, reusable Instructions carry your voice and product context into every new prompt, and the free Chrome extension puts the library inside ChatGPT, Claude, and Gemini. The prompt library walkthrough shows the screens.
Pricing: free plan with 5 prompts, no credit card. Starter is $9 a month (1,000 credits), Pro is $19 a month (3,000 credits), and Unlimited is $49 a month (fair-use limits after 12,000 credits). Yearly billing saves 20%. See pricing.
Best for: marketers, founders, writers, and support teams who used OpenAI's optimizer to clean up prompts for ChatGPT. Start with the ChatGPT prompt generator or the Claude prompt generator if you are moving work to Claude.
Skip it if: you need graded runs over a dataset or prompts served to your app through an API. Prompt Builder has no runtime API and does not score outputs against graders.
2. Promptfoo
Promptfoo is the replacement OpenAI itself points to. It is an open-source CLI and library for evaluating and red-teaming LLM apps, MIT licensed, with about 27,300 GitHub stars, and its README now says "Promptfoo is now part of OpenAI." You describe prompts, providers, test cases, and assertions in a config file, then run promptfoo eval and open the results with promptfoo view, locally or in CI.
The mapping from OpenAI Evals is direct: your test data becomes test cases, your graders become assertions and metrics, and your runs move from the dashboard into your repo. OpenAI's guide warns that similarity scores may not match exactly across systems, so re-check any LLM-as-a-judge grader before trusting it for regression decisions.
Pricing: Community is free forever and includes up to 10,000 red-teaming probes a month. Enterprise (team sharing, SSO, managed cloud) and On-Premise are custom priced.
Best for: teams whose real dependency was Evals and graders, not the Optimize button.
Skip it if: you want the prompt rewritten for you. Promptfoo tells you which prompt wins; it does not write the new one.
3. DSPy
DSPy is the closest open-source match for what the dataset-backed optimizer did. It is a Python framework (MIT license) where you describe a task as a signature, give it training examples and a scoring function, and let an optimizer tune the prompts. Named optimizers include GEPA (reflective prompt evolution), MIPROv2, BootstrapFewShot, COPRO, and SIMBA.
The DSPy homepage shows a GEPA run lifting an F1 score from 0.41 to 0.63, and a second example taking accuracy from 62% to 89% on 200 examples. Your results depend on your metric, the same rule OpenAI's docs gave for graders.
Pricing: free. You pay your model provider for the optimization calls, which can be many.
Best for: developers with labeled examples and a metric who want the optimizer loop inside their own pipeline, on any model.
Skip it if: nobody on your team writes Python, or you have no examples to score against.
4. Vertex AI Prompt Optimizer
Google's optimizer has two modes. The zero-shot optimizer is a real-time, low-latency tool that improves a single prompt or system instruction with no examples needed. The data-driven optimizer is a batch, task-level optimizer that improves prompts by evaluating the model's responses to a few labeled samples, which is the same shape as OpenAI's dataset-backed tool.
Google says the optimizer is generally available, while its SDK library is still experimental. The data-driven mode supports generally available Gemini models plus custom models you deploy, and the zero-shot mode has a setting for small models such as Gemini Nano and Gemma 3n.
Pricing: the docs list no separate fee for the optimizer; it runs in your Google Cloud project, so expect normal model charges for the calls it makes.
Best for: teams already on Google Cloud, or teams moving production prompts from GPT to Gemini.
Skip it if: your app stays on OpenAI models and you do not want a second cloud account.
5. Braintrust
Braintrust is a hosted eval and observability platform. Its Loop agent can run evaluations, generate test cases, and iterate on prompts on its own, which is the nearest hosted equivalent to "annotate, optimize, repeat" in the OpenAI dashboard.
Pricing: Starter is $0 with unlimited users, $10 of included model credits, 1 GB of processed data, 10,000 scores, and 14-day data retention, no credit card required. Pro is $249 a month with $100 of model credits, 5 GB, 50,000 scores, and 30-day retention. Starter allows 5 concurrent Loop threads without on-demand usage.
Best for: product teams that liked having datasets, runs, and prompt edits in one web UI and do not want to live in a YAML file.
Skip it if: you are one person cleaning up chat prompts. The jump from free to $249 a month is built for teams shipping an LLM feature.
6. PromptLayer
PromptLayer is a prompt management platform: a registry where prompts are versioned and fetched by your app, plus evaluations. That makes it a direct answer to the other November 30 shutdown, reusable prompt objects, if you want prompts out of your code and editable by non-engineers. We compared it with other registries in our prompt manager roundup.
Pricing: Free covers 10 prompts, 5 users, 2,500 requests, and 250 eval cell executions a month. Pro is $49 a month with unlimited prompts. Team is $500 a month for 25 users. Enterprise is custom.
Best for: teams that called pmpt_ IDs from code and want the same "edit the prompt without a deploy" workflow.
Skip it if: you only need prompts in a file. OpenAI's own advice, prompts in your codebase reviewed like code, is free.
7. Prompt Optimizer (open source, by linshenkx)
The most popular open-source rewriter, with about 41,700 GitHub stars. It runs as a web app at prompt.always200.com (data stays in your browser), a desktop app, a Chrome extension, or a Docker container. It works with OpenAI, Gemini, DeepSeek, Grok, and any OpenAI-compatible API, and it has analysis, single-result evaluation, and multi-result compare evaluation, so you can check whether a rewrite actually helped.
Pricing: free under AGPL-3.0. You pay your model provider per token.
Best for: developers and privacy-minded users who want a free rewriter and are fine handling their own API key.
Skip it if: you want a saved, shared library or do not want to manage keys.
How to choose
- You clicked Optimize on single prompts for ChatGPT and never built a dataset: use a rewriter. Prompt Builder if you want it hosted with a library, the open-source Prompt Optimizer if you want it free with your own key. Our prompt optimizer roundup compares 8 of them.
- You relied on Evals and graders: move the tests to Promptfoo first, since that is OpenAI's supported path, then add DSPy if you want automatic rewriting against the same metric.
- You want a hosted UI for datasets, runs, and prompt iteration: Braintrust.
- You are moving to Gemini anyway: Vertex AI's data-driven optimizer.
- You called reusable prompts by ID: put the prompt text in your code (free), or use PromptLayer if non-engineers need to edit prompts without a deploy.
A migration checklist for the next two weeks
- List every prompt you saved in the platform. Copy the final text and the version you actually use into your repo or prompt library.
- Copy your dataset rows and annotations. Annotations with written feedback are the expensive part to rebuild; the model outputs can be regenerated.
- Write down each grader as a plain rule. "Must mention the refund window," "under 120 words," "valid JSON with these keys." Those rules become Promptfoo assertions or a DSPy metric.
- Replace every
prompt: { id: ... }call with inline instructions before November 30, then test the endpoint. - Run the old and new prompt side by side once. In our prompt structure study, structured prompts scored 18.97 out of 20 against 9.56 for a bare one-line ask, so a well-structured prompt carries over between models better than most people expect, but check it on your own inputs.
FAQ
When does the OpenAI prompt optimizer shut down? OpenAI's dataset-backed prompt optimizer is part of the Evals platform, which went read-only for existing users on October 31, 2026 and shuts down on November 30, 2026.
What does OpenAI recommend instead of Evals? Promptfoo. OpenAI's deprecations page links a migration guide from OpenAI Evals to Promptfoo, an open-source eval CLI that is now part of OpenAI. You recreate each eval by hand in a config file.
Is there a free alternative to the OpenAI prompt optimizer? Yes. DSPy and Promptfoo are free and open source, the open-source Prompt Optimizer is free with your own API key, Braintrust has a $0 Starter plan, and Prompt Builder has a free plan with 5 prompts.
What replaces OpenAI reusable prompts (/v1/prompts)? OpenAI says to move the prompt content into your application code. If non-engineers need to edit prompts without a deploy, a prompt registry such as PromptLayer does the same job.
Does Promptfoo optimize prompts? No. Promptfoo runs prompts against test cases and scores the outputs. Pair it with DSPy, Vertex AI's optimizer, or a rewriter like Prompt Builder to produce the new prompt, then use Promptfoo to confirm it is better.
Can I still use OpenAI Evals before November 30? Existing users can read their evals and results, but the platform has been read-only since October 31, 2026. Copy what you need before the shutdown date.
Bottom line
The OpenAI prompt optimizer was three tools in one dashboard: a rewriter, a scorer, and a place to keep prompts. After November 30, pick one replacement per job. Promptfoo covers scoring and is OpenAI's own path. DSPy or Vertex AI covers dataset-driven rewriting. For prompts you write by hand for ChatGPT, Claude, or Gemini, try the AI prompt generator on your current prompt and keep the winners in your prompt library.